ttaoretw / tacotron-pytorch Goto Github PK

Hello!
Have you think about developper a code for export this model to coffee2 or anything for use in app android?

Thanks!

Exactly what the title says - how do I use the pretrained model?

模型的输入是 Word Embedding 还是 Character Embedding ？ 。 “TACOTRON: TOWARDS END-TO-END SPEECH SYN- THESIS” 论文里面输入的 Character Embedding 。 我大概看了一下代码好像找不到关于Character Embedding 的部分

Hello!
I'm trying training my spanish voice, but i have problems with especials characters (for example: ñ)

But i try my own wavs files and metadata without any special character but i have the same problem, when i strart training (main.py) i have this error:

[INFO] Load data
[INFO] Build model
Traceback (most recent call last):
File "main.py", line 36, in
solver.build_model()
File "/content/Tacotron-pytorch/src/solver.py", line 80, in build_model
self.model = Tacotron(**self.config['model']['tacotron']).to(device=self.device)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 386, in to
return self._apply(convert)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 193, in _apply
module._apply(fn)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 193, in _apply
module._apply(fn)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 193, in _apply
module._apply(fn)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/rnn.py", line 127, in _apply
self.flatten_parameters()
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/rnn.py", line 123, in flatten_parameters
self.batch_first, bool(self.bidirectional))
RuntimeError: cuDNN error: CUDNN_STATUS_EXECUTION_FAILED

I'm using default config:
tacotron:
n_vocab: 250 # Use a larger vocab size is ok.
embedding_size: 256
mel_size: 80
linear_size: 1025
r: 5

The self.device it's correct because with LJSpeech work fine.
Have you any idea about this error?

Thanks man!
PD: excellents result with LJSpeech

Hello!
I see that you don't save the .pth if linear_loss is bigger than the rest of steeps. I think is good idea but i use the google colab (TeslaGPU) for training and i only have sessions of 12hours.

When i use --checkpoint-path i can see that the best-linear-loss start in 1e10 and i created new param for load last best-linear-loss, you can see:

is normal in step 16K linear_loss: 0.042?
My problem is that i can't give less than 0.042 in 12 hours and i don't know if is good save step 40K even though the linear_loss is bigger?

Thanks for your help!
Good code!

Hi tao,
Thank you for your effort to code the Tacotron. I would like to ask why do we get different results in tmp/results and eval/results even if we use the same sentences for eval ?
The train results in tmp sounds very good, but in eval, it sounds like it doesn't work. Shouldn't it be the same if we use same sentences ?
Thanks in advance

I have a model with 100k in spanish and work ok.

Need make any change in config for continue this .pth with news wavs for make other people voice?

thanks!

def txt2seq(txt):
txt = ''.join([t for t in txt if t in _characters])
txt = english_cleaners(txt) + _eos
seq = [symbol2id[s] for s in txt]
return seq

I guess 'english_cleaners' function should be called ahead because first line erases numbers.